save-context

Warn

Audited by Gen Agent Trust Hub on Mar 23, 2026

Risk Level: MEDIUMDATA_EXFILTRATIONCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [DATA_EXFILTRATION]: The skill automatically performs a git push of the generated context snapshot if the storage directory is identified as a git repository. This behavior transmits potentially sensitive conversation data (including goals, decisions, and file paths) to a remote server without a separate user confirmation step for the network operation.
  • [COMMAND_EXECUTION]: The skill executes shell commands (git add, git commit, git push, cd) to manage the saved context files. It utilizes the $ARGUMENTS variable to define filenames without explicit sanitization, which presents a surface for command injection if the platform does not provide inherent protection.
  • [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection through the processing of conversation history.
  • Ingestion points: Full conversation history (SKILL.md).
  • Boundary markers: Absent; there are no delimiters or instructions to ignore embedded commands within the conversation history being analyzed.
  • Capability inventory: File system writes to multiple directories, shell command execution, and network transmission via git push.
  • Sanitization: Absent; the content extracted from the conversation is interpolated directly into the markdown output template.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 23, 2026, 10:58 AM
Security Audit — agent-trust-hub — save-context